Skip to content
NewsModel release

AWS open-sources Strands Decider 2B, a decision model that answers in 115 milliseconds

· by Pondero Newsdesk

The short version

AWS released a 2-billion-parameter open model that scores a developer's fixed set of answers instead of generating text, landing days after OpenAI shipped a hosted rival for the same new model category.

AWS open-sources Strands Decider 2B, a decision model that answers in 115 milliseconds

AWS's Strands Labs released Strands Decider 2B on October 1, an open source model that does not write text at all. It picks between a developer's predefined options and attaches a confidence score, running locally in a median of about 115 milliseconds, per AWS's announcement.

What

Decider strips the text-generating head off a pretrained "torso," in this case Qwen3.5-2B, and replaces it with a pointer head of just over a million parameters that scores each supplied option against the model's hidden state, according to AWS's blog post. The torso is fine-tuned with a rank-16 LoRA adapter. AWS says the result is its first contribution to a new class the company calls "decision models" or "system one models": smaller, faster checkpoints meant to pick an answer reliably rather than generate open-ended output. On a local Nvidia RTX 3090, AWS reports a median decision latency around 115 milliseconds, rising for larger tasks; on an M3 MacBook, small tasks land around 153 milliseconds, per the same post. AWS also says the 2B model ranks third of 33 models in its size class on JevBench's public accuracy-and-calibration set, first of 30 once models just over 2 billion parameters are excluded, and answers every one of JevBench's easy-tier questions correctly. The weights are on Hugging Face and the code, training data, and training scripts are on GitHub, per AWS.

AWS ships the open alternative to OpenAI's hosted Decisions API

Decider landed two days after OpenAI's own entry into the same category: a Decisions API built on a model called Luna, launched Tuesday, September 30 as a limited preview, according to The New Stack. The split matters for anyone building agent workflows now: OpenAI's version is a hosted API, while AWS shipped a downloadable checkpoint with the full training recipe included, so a team can inspect it, retrain it, or run it entirely offline instead of depending on a vendor endpoint. Both are chasing the lead of TypeSafe's Jev, which AWS says kicked off the current wave of decision models when it briefly topped JevBench's leaderboard earlier this year. AWS distinguished engineer Marc Brooker told TechCrunch that AWS customers wanted "a workflow step that can be structured in a way that is more reliable...lower latency, potentially lower cost" than routing every agent decision through a full LLM call, per TechCrunch. TypeSafe's CEO, Diogo Almeida, was less impressed by the rush of clones: "I get that people think it's a gold rush, but they might be underestimating the difficulty of making the models actually smart," he told TechCrunch.

For teams already building on Strands, AWS's own demo shows the practical use case: an agent that guesses which city a user meant before calling a weather tool gets checked by Decider first, and gets sent back to ask for clarification if the guess is not grounded in the conversation, per AWS. That is a cheap, local gate in front of an action, not a replacement for the agent's main model.

What to watch next

AWS has not said whether Decider gets a hosted option inside Bedrock, which The New Stack flags as the open question for teams that want this pattern in production rather than on a developer's laptop. Also worth tracking is whether JevBench, already the shared yardstick AWS, TypeSafe, and independent projects like Kev and Laya are citing, becomes the field's default benchmark as more vendors enter the category.

Sources